AI Art and its Impact on Artists

Harry Jiang Lauren Brown Jessica Cheng

Independent Researcher Artist Artist
Canada USA Canada
hhj@alumni.cmu.edu labillustration@gmail.com chengelingart@gmail.com

Anonymous Artist Mehtab Khan Abhishek Gupta

Artist Yale Law Montréal AI Ethics Institute
USA USA Canada
mehtab.khan@yale.edu abhishek@montrealethics.ai

Deja Workman Alex Hanna Jonathan Flowers

Penn State University The Distributed AI Research Institute California State University,
USA USA Northridge
dqw5409@psu.edu alex@dair-institute.org USA
johnathan.flowers@csun.edu

Timnit Gebru
The Distributed AI Research Institute
USA
timnit@dair-institute.org

ABSTRACT
The last 3 years have resulted in machine learning (ML)-based
image generators with the ability to output consistently higher
quality images based on natural language prompts as inputs. As
a result, many popular commercial “generative AI Art” products
have entered the market, making generative AI an estimated $48B
industry [125]. However, many professional artists have spoken
up about the harms they have experienced due to the proliferation
of large scale image generators trained on image/text pairs from
the Internet. In this paper, we review some of these harms which
include reputational damage, economic loss, plagiarism and copy-
right infringement. To guard against these issues while reaping the
potential benefits of image generators, we provide recommenda-
tions such as regulation that forces organizations to disclose their
training data, and tools that help artists prevent using their content
as training data without their consent.
ACM Reference Format:
Harry Jiang, Lauren Brown, Jessica Cheng, Anonymous Artist, Mehtab Khan,
Abhishek Gupta, Deja Workman, Alex Hanna, Jonathan Flowers, and Timnit
Gebru. 2023. AI Art and its Impact on Artists. In AAAI/ACM Conference on
AI, Ethics, and Society (AIES ’23), August 08–10, 2023, Montréal, QC, Canada.
ACM, New York, NY, USA, 12 pages. https://doi.org/10.1145/3600211.3604681
This work is licensed under a Creative Commons Attribution International
4.0 License.
AIES ’23, August 08–10, 2023, Montréal, QC, Canada
© 2023 Copyright held by the owner/author(s).
ACM ISBN 979-8-4007-0231-0/23/08.
https://doi.org/10.1145/3600211.3604681
1 INTRODUCTION
In the two years since the publication of [18] which outlines the
dangers of large language models (LLMs), multimodal generative
artificial intelligence (AI) systems with text, images, videos, voice,
and music as inputs and/or outputs have quickly proliferated into
the mainstream, making the generative AI industry valued at an es-
timated $48B [125]. Tools like Midjourney [78], Stable Diffusion [5],
and DALL-E [91] that take in text as input and output images, as
well as image-to-image based tools like Lensa [97] which output
altered versions of the input images, have tens of millions of daily
users [47, 127]. However, while these products have captured the
public’s imagination, arguably to a much larger extent than any
prior AI system, they have also resulted in tangible harm, with
more to come if the ethical concerns they posit are not addressed
now. In this paper, we outline some of these concerns, focusing
our discussion on the impact of image based generative AI systems,
i.e. tools that take text, images, or a combination of both text and
images as inputs, and output images. While other works have sum-
marized some of the potential harms of generative AI systems more
generally [18, 28, 29], we focus our discussion on the impacts of
these systems on the art community, which has arguably been one
of the biggest casualties (Section 4) [40, 138].
As we argue in Section 3, image based generative AI systems,
which we call image generators throughout this paper, are not
artists. We make this argument by first establishing that art is a
uniquely human endeavor, using perspectives from philosophies
of art and aesthetics. We further discuss how anthropomorphizing
image generators and describing them as merely being “inspired”
by their training data, like artists are inspired by other artists, is
not only misguided but also harmful. Ascribing agency to image
generators diminishes the complexity of human creativity, robs
artists of credit (and in many cases compensation), and transfers

363
AIES ’23, August 08–10, 2023, Montréal, QC, Canada
accountability from the organizations creating image generators,
and the practices of these organizations which should be scrutinized,
to the image generators themselves.
While companies like Midjourney, Stability AI and Open AI
who produce image generators are valued at billions of dollars
and are raising hundreds of millions of dollars1 , their products are
flooding the market with content that is being used to compete with
and displace artists. In section 4, we discuss the impact of these
products on working artists, including the chilling effect on cultural
production and consumption as a whole. Merely open sourcing
image generators does not solve these problems as they would still
enable people to plagiarize artists’ works, and impersonate their
style for uses that the artists have not consented to.
In Section 5, we provide a summary of the relevant legal ques-
tions pertaining to image generators. While there have been legal
developments around the world, we focus our analysis on the US
where a number of lawsuits have been filed by artists challenging
the use of image generators [129]. Given that copyright has been
the most frequently invoked law in such cases [28], we provide an
overview discussing the relevance of US copyright law in protecting
artists, and conclude that it is largely unequipped to tackle many
of the types of harms posed by these systems to content creators.
As we discuss in Section 6, the AI research community has enabled
the aforementioned harms through data laundering, with for-profit
corporations partnering with academic institutions that help them
gather training data for commercial purposes while increasing their
chances of courts finding these uses to be “fair use”.
We end our discussion with proposals for new tools and regula-
tions that tackle some of the harms discussed in this paper, as well
as encouraging the AI community to align themselves with those
harmed by these systems rather than powerful entities driving the
proliferation of generative AI models trained on the free labor of
content creators.
2 LITERATURE REVIEW
2.1 Background on Image Generation
We define “generative artificial intelligence (AI)” to encompass ma-
chine learning products that feature models whose output spaces
overlap in part or in full with their input spaces during training,
though not necessarily inference. While generative AI systems
are based on generative models which statistically aim to model
the joint distribution between a feature space and output space
𝑝 (𝑥, 𝑦) [85], we distinguish between “generative AI” systems and
generative models as the latter can be used in classification sys-
tems. This paper focuses on products whose stated output space
composes, in part or in full, of visual data (i.e. images), which will
be referred to as image generators; similarly, the scope of art dis-
cussed within this work is largely limited to the fields of visual art.
We consider two different applications in the context of inference,
text-to-image and image-to-image, though more recent multimodal
pretrained model architectures usually are capable of both (and
often necessitate both).
Early approaches to image synthesis such as [38, 95, 120], aimed
to achieve texture synthesis, i.e. modifying an existing image to
1 https://www.nytimes.com/2023/01/23/business/microsof t- chatgpt- artificial-
intelligence.html
364
copy the texture of another image [38, 95, 120]. In the deep learning
era of computer vision (2012 until now), Convolutional Neural
Networks (CNNs) enabled the ability to recognize a large amount
of latent attributes that do not conform to arbitrary statistical forms,
unlike early works in texture synthesis [69].
In addition to CNNs, another architectural element of note is
the variational autoencoder (VAE), like the one used in Yan et
al. [135]; VAEs, which use two mirrored neural network compo-
nents to map the input space to a latent space (encoder) and vice
versa (decoder) [68], set the stage for the development of gener-
ative models, which significantly widened the capacity of image
synthesis. A key element of VAEs is the reconstructive loss function
which allows an ML system to explicitly define its training objec-
tive as the re-creation of input features, with the expectation that
the model can generalize beyond the training set during inference.
VAEs enabled the creation of image generation models such as
VQ-VAE-2 [104] and are components of many subsequent models.
The next major breakthrough is the generative adversarial net-
work (GAN), which employs the use of two models trained simul-
taneously [58]. Unlike conventional neural networks such as VAEs
which directly and asymmetrically measure the divergence between
a distribution known to be a reference and one known to be a hy-
pothesis, GANs indirectly measure the divergence between two
distributions of masked origin through the intermediary of the
discriminator. The introduction of conditional losses in [82] made
GANs the dominant architecture in image generation due to the
ability to now inform outputs with text tags as auxiliary informa-
tion; the paper itself used a handwriting generator trained on the
MNIST dataset [35] as a demonstration. With GANs came the first
large-scale image generating models, allowing for output sizes of
up to 512 × 512 [24, 61, 65].
The adaptation of approaches from natural language processing
(NLP) such as transformers further enabled having complex text as
input for text-to-image models [42]. In [30], OpenAI adapted the
architecture of GPT-2, a large language model (LLM), to output a
series of pixel values that could be rearranged into a recognizable
image. This research led to the original DALL-E [103], a tool that
outputs a 256 × 256 RGB image based on natural language prompts,
this time using the GPT-3 architecture [25].
In the last 3 years, the use of GANs for image generation has
been overtaken by diffusion models which take inspiration from
fluid dynamics [37, 86, 116, 118]. These models work by repeatedly
applying gaussian noise on an image (imitating the diffusion process
of fluids or heat), and then denoising the result in equally many
steps [118]. In a departure from GANs’ implicit modeling, diffusion
models return to using a reconstruction loss.
In 2022, Rombach et al. released the Stable Diffusion model [4,
106], which uses a conditional latent space based on text and im-
ages: in this case a pretrained model by OpenAI called CLIP [99].
This allowed for models that are not confined to natural language
understanding (NLU)-based architectures, and can generate high-
quality images based on natural language prompts. In the same
year, OpenAI released DALL-E 2 [91] which has a similar model
architecture [102] but with a training dataset that is opaque to the
public.
AI Art and its Impact on Artists
In addition to different model architectures, massive image datasets
such as JFT-300M (300M images) [124] have helped improve im-
age generation performance. The current crop of image genera-
tors, primarily those based on Stable Diffusion, are pretrained on
LAION [109], or its variants which are subsets of the original 5B
dataset. The dataset consists of 5.85 billion CLIP-filtered image-text
pairs, of which 2.32B contain English language text. An exploration
of a subset of LAION can be found at [11].
2.2 Products for Image Generation
The advent of Stable Diffusion and related models has resulted in
a proliferation of commercial and non commercial image genera-
tion tools that use them. Stability AI’s Stable Diffusion [5] and its
commercial product Dream Studio2 , OpenAI’s DALL-E 2 [91], and
Midjourney [78] are the most popular systems built on diffusion
models, with StarryAI [122], Hotpot.ai [94], NightCafe [123], and
Imagen [108] being a few others. Established art software company
Adobe has also released its image generator product, Adobe Fire-
fly [3], which the company says is trained on Adobe Stock images,
images in the public domain, and those under open licensing. The
ecosystem is large and expanding, including organizations like Fo-
tor [45], Dream by WOMBO [133], Images.AI [128], Craiyon [71],
ArtBreeder [9], Photosonic [134], Deep Dream Generator [55], Run-
way ML [107], CFSpark [46], MyHeritage Time Machine [73], and
Lensa [97]. While some advertise the model architectures they use,
such as StableCog [121] using diffusion-based techniques, others
provide little to no detail. For example, while the CEO of Stability AI
has written that Midjourney used Stable Diffusion in past releases 3 ,
Midjourney does not disclose underlying model information for
its current releases, only mentioning “a brand-new AI architecture
designed by Midjourney” in describing its releases since November
2022 [79].
Most of the products identified above emerged as specific com-
mercial offerings for users to generate images by providing text
prompts. There are other services that have been introduced as fea-
tures in existing products, such as synthetic images in Canva [26],
Shutterstock [113], and Adobe Stock Images [2], which seek to
augment their stock image offerings with synthetic images. On the
other hand, companies like Getty Images took a stance against in-
cluding synthetic images in their portfolio of offerings in 2022 [130],
although NVIDIA announced a collaboration with them in 2023
to develop image generators [76]. Open source efforts in the space
have focused on using Stable Diffusion and other open-source vari-
ants to create plugins for Photoshop [7], Unreal Engine [43], and
GIMP [20]. Some groups, such as Unstable Diffusion, are explicitly
focused on generating not-safe-for-work (NSFW) content [59].
3 IMAGE GENERATORS ARE NOT ARTISTS
Many researchers have pointed out the issues that arise from the
anthropomorphization of AI systems, including shifting responsi-
bility from the people and organizations that build these systems,
to the artifacts they build as if those artifacts have agency on their
own [13, 16, 39]. This anthropomorphization is readily apparent
2 https://dreamstudio.com
3 https://web.archive.org/web/20220823032632/https://twitter.com/EMostaque/status
/1561917541743841280, referring to V3
365
AIES ’23, August 08–10, 2023, Montréal, QC, Canada
in descriptions of image generators as if they are artists [39], even
going as far as to claim that the image generators are “inspired”
by the data they are trained on, similar to how artists are inspired
by other artists’ works [66]. In this section, we discuss why such
arguments are misguided and harmful.
Following philosophers of art and aesthetics from varied dis-
ciplines (e.g. Chinese and Japanese Philosophy, American Prag-
matism, and Africana Philosophy), we define art as a uniquely
human endeavor connected specifically to human culture and ex-
perience [6, 36, 62, 74, 75, 88, 93]. Most philosophers of art and
aesthetics argue that while non-human entities can have aesthetic
experiences and express affect, a work of art is a cultural product
that uses the resources of a culture to embody that experience in a
form that all who stand before it can see. On this view, art refers to
a process that makes use of external materials or the body to make
present experience in an intensified form. Further, this process
must be controlled by a sensitivity to the attitude of the perceiver
insofar as the product is intended to be enjoyed by an audience.
The artwork, therefore, is the result of a process that is controlled
for some end and is not simply the result of a spontaneous activity
( [36] pp. 54, 55). This control over the process of production is what
marks the unique contribution of humanity: while art is grounded
in the very activities of living, it is the human recognition of cause
and effect that transforms activities once performed under organic
pressures into activities done for the sake of eliciting some response
from a viewer. As an example, a robin might sing, a peacock might
dance, but these things are performed under the organic pressures
of seeking a mate. In humans, song and dance are disconnected from
the organic pressures of life and serve purposes beyond the mere
satisfaction and expression of organic pressures, and serve cultural
purposes. In brief, art is a form of communication: it communicates.
In contrast, the outputs of artifacts like image generators are not
framed for enjoyment because they merely imitate the technical
process, and then only those technical processes embodied in the
works that make up the training dataset. The image generator
has no understanding of the perspective of the audience or the
experience that the output is intended to communicate to this
audience. At best, the output of image generators is aesthetic, in
that it can be appreciated or enjoyed, but it is not artistic or art
itself. Thus, “Mere perfection in execution, judged in its own terms
in isolation, can probably be attained better by a machine than by
human art. By itself, it is at most technique. . . To be truly artistic, a
work must also be esthetic—that is, framed for enjoyed receptive
perception.” ( [36] pp. 54).
Thus, art is a uniquely human activity, as opposed to something
that can be done by an artifact. While image generators have to be
trained by repeatedly being shown the “right” output, using many
examples of the desired target, and explicitly defining an objective
function over which to optimize, humans do not have such rigid
instructions. In fact, while image generators have been shown to
even memorize their data and can output almost exact replicas of
images from their training set under certain conditions [27, 117],
as artist Karla Ortiz writes, artists’ styles are so unique to them,
that it is very difficult for one artist to copy another’s work [92].
The very few artists who are able to do this copying are known for
this skill [92]. An artists’ ‘personal style’ is like their handwriting,
authentic to them, and they develop this style (their personal voice
AIES ’23, August 08–10, 2023, Montréal, QC, Canada
and unique visual language) over years and through their lived
experiences [92].
The adoption of any particular style of art, personal or other-
wise, is a result of the ways in which the individual is in transaction
with their cultural environment such that they take up the customs,
beliefs, meanings, and habits, including those habits of aesthetic
production, supplied by the larger culture. As philosopher John
Dewey argues, an artistic style is developed through interaction
with a cultural environment rather than bare mimicry or extrap-
olation from direct examples supplied by a data set [23]. Steven
Zapata argues, “our art ‘creates’ us as artists as much as we create
it” [138]. This experience is unique to each human being by virtue
of the different cultural environments that furnish the broader set
of habits, dispositions towards action, that enabled the development
of anything called a personal style through how an individual took
up those habits and deployed them intelligently.
Finally, an image generator is trained to generate images from
prompts by mapping images and texts into a lower dimensional
representation in a latent space [58, 68, 106]. This latent space is
learned during the model’s training process. Once the model is
trained, this latent space is fixed and can only change through train-
ing from scratch or fine-tuning on additional examples of image-text
pairs [57]. In contrast, human inspiration changes continuously
with new experiences, and a human’s relationship with their lived
experiences evolves over time. Most importantly, these experiences
are not limited to additional artistic training or viewing of images.
Rather, humans perform abstract interpretations between repre-
sentational and imaginary subjects, topics, and of course, personal
feelings and experiences that an artifact cannot have.
Let’s look at Katsuhiro Otomo’s seminal Akira as an example.
Otomo notes that he created these images by drawing inspiration
from his own teenage years, thinking about a rebuilding world,
foreign political influence, and an uncertain future after World War
II [12]. Similarly, Claude Monet created his defining Nymphéas
[Water Lilies] series during the last 30 years of his life, after the
loss of his son in 1914 [63]. As shown by both these artists, and
many other artists, the human experience both defines and inspires
creation across an artist’s personal lifetime. Each individual’s art is
unique to their life experiences. Otomo’s Akira is a fundamentally
different form of artwork than Monet’s Nymphéas [Water Lilies]
series not simply due to their different stylistic and pictorial media,
but due to the way in which each artists’ work was an expression
of a cultural inheritance that shaped the unique experiences that
gave rise to their particular art forms. While image generators can
imitate the stylistic habits, the “unique voices” of a given artist, they
cannot develop their own particular styles because they lack the
kinds of experiences and cultural inheritances that structure every
creative act. Even when provided with a human-written prompt, the
sampling of a probability distribution conditional on a string of text
does not present a synthesis of concepts, emotion, and experience.
In conclusion, image generators are not artists: they require
human aims and purposes to direct their “production” or “repro-
duction,” and it is these human aims and purposes that shape the
directions to which their outputs are produced. However, many
people describe image generators as if these artifacts themselves
are artists, which devalues artists’ works, robs them of credit and
366
compensation, and ascribes accountability to the image genera-
tors rather than holding the entities that create them accountable.
In [39], Epstein et al. performed a study with participants on Ama-
zon Mechanical Turk to assess the impact of anthropomorphization
of image generators, finding a relationship between the manner in
which participants assign credit and accountability to stakeholders
involved in training and producing image generators, and the level
of anthropomorphization. They advise “artists, computer scientists,
and the media at large to be aware of the power of their words, and
for the public to be discerning in the narratives they consume.”
4 IMPACT ON ARTISTS
The proliferation of image generators poses a number of harms to
artists, chief among them being economic loss due to corporations
aiming to automate them away. In this section, we summarize some
of these harms, including the impact of artists’ styles being mim-
icked without their consent, and in some cases, used for nefarious
purposes. We close with a discussion of how image generators stand
to perpetuate hegemonic views and stereotyping in the creative
world, and the chilling effects of these technologies on artists as
well as overall cultural production and consumption.
4.1 Economic Loss
While artists hone their craft over years of practice, observation,
and schooling, having to spend time and resources to pay for sup-
plies, books, and tutorials, companies like Stability AI are using
their works without compensation while raising billions from ven-
ture capitalists to compete with them in the same market4 . Leaders
of companies like Open AI and Stability AI have openly stated
that they expect generative AI systems to replace creatives immi-
nently56 . Stability AI CEO Emad Mosque has even accused artists of
wanting to have a “monopoly on visual communications” and “skill
segregation”7 . To the contrary, current image generation business
models like those of Midjourney, Open AI and Stability AI, stand
to centralize power in the hands of a few corporations located in
Western nations, while disenfranchising artists around the world.
It is now possible for anyone to create hundreds of images in
minutes, compile a children’s book in an hour8 , and a project for a
successful Kickstarter campaign in a fraction of the time it takes for
an actual artist9 . Although many of these images do not have the
full depth of expression of a human, commercial image generators
flood the market with acceptable imagery that can supplant the
demand for artists in practice. This has already resulted in job
losses for artists, with companies like Netflix Japan using image
generators for animation, blaming “labor shortage” in the anime
industry for not hiring artists [32].
4 https://techcrunch.com/2022/10/17/stability- ai- the- startup- behind- stable-
diffusion-raises-101m/
5 https://web.archive.org/web/20220912045000/https://twitter.com/sama/statu
s/1484950632331034625, https://web.archive.org/web/20220122181741/https:
//twitter.com/sama/status/1484952151222722562
6 https://web.archive.org/web/20230811193157/https://twitter.com/emostaque/status
/1591436813750906882
7 https://web.archive.org/web/20230224175654/https://twitter.com/mollycrabapple/s
tatus/1606148326814089217
8 https://www.youtube.com/watch?app=desktop&v=ZbVRYqsntDY
9 https://web.archive.org/web/20230124003305/https://twitter.com/spiridude/status/1
616476006444826625
AI Art and its Impact on Artists
One of the more high profile cases of the labor impact can be
seen in the title sequence of Marvel Studio’s 2023 TV series Secret
Invasion, which uses a montage of generated imagery [81]. While
prior movies from the studio feature between 5 (The She-Hulk:
Attorney at Law 10 ) and 9 (Hawkeye11 ) artists and illustrators for
their title sequences, Secret Invasion has only one “Sagans Carle”
credited as “AI Technical Director”12 . This labor displacement is
evident across creative industries. For instance, according to an
article on Rest of World, a Chinese gaming industry recruiter has
noticed a 70% drop in illustrator jobs, in part due to the widespread
use of image generators [139]; another studio in China is reported
to have laid off a third of its character design illustrators [139].
In addition to displacing the jobs of studio artists, the noise
caused by the amount of AI-generated content will likely be dev-
astating for self-employed artists in particular. This has become
evident in the literary world with the advent of LLM based tools
like ChatGPT13 . Recently, Clarkesworld, a popular science fiction
magazine, temporarily closed open submissions after being over-
whelmed by the number of ChatGPT generated submissions they
received [31]. They announced that they will instead only solicit
works from known authors, which disadvantages writers who are
not already well known. It is not difficult to extrapolate such a
result with visual art venues that receive too many AI-generated
images. Contrary to “democratizing art,” this reduces the number
of artists who can share their works and receive recognition.
Regardless of their objections, some working artists have started
to report having to use image generators to avoid losing their
jobs, further normalizing its commercial use [139]. Artists have
also reported being approached by companies producing image
generators to work on modifying the outputs of their systems 14 .
This type of work reduces hard earned years of skill and artistic eye
to simple cleanup work, with no agency for creative decisions. In
spite of these issues, creatives in executive roles who can be isolated
from the realities of most working artists, may gravitate towards
using these tools without considering the effects on the industry at
large, such as a reduction in the economic earning power of many
working artists. For instance, the director of Secret Invasion had
editorial control in deciding whether to use image generators15 , and
chose to replace illustrators’ works with image generated content.
With the increasing barriers and job losses for creatives because
of image generators, the pursuit of art could be relegated to the inde-
pendently wealthy and those who can afford to develop their artistic
skills while working a full-time job. This will disproportionately
harm the development of artists from marginalized communities,
like disabled artists, and artists with dependents.
4.2 Digital Artwork Forgery
As discussed in Section 2, image generators are trained using bil-
lions of image-text pairs obtained from the Internet. Stable Diffusion
V2, for instance, is trained using the publicly available LAION-5B
10 https://ondisneyplus.disney.com/show/she-hulk
11 https://ondisneyplus.disney.com/show/hawkeye
12 https://www.disneyplus.com/series/invasion-secreta/3iHQtD1BDpgN
13 https://openai.com/blog/chatgpt
14 https://www.facebook.com/story.php?story_fbid=pfbid02L9Qkj6Bnidy6zL7hRjv
Q9MuYLQF3jSUXcGLRjjgZhxH1LysnV4DZRUgMyhLMvKxGl&id=882110175
15 https://www.polygon.com/23767640/ai-mcu-secret-invasion-opening-credits
367
AIES ’23, August 08–10, 2023, Montréal, QC, Canada
dataset [106, 109]. Although the creators of LAION-5B have not
provided a way for people to browse the dataset, various artists
have reported finding their works in the training data without their
consent or attribution [11]. Open AI has not shared the dataset
that its image generator, DALL-E, was trained on, making it im-
possible to know the extent to which their training data contains
copyright protected images. Using a tool16 built by Simon Willi-
son which allowed people to search 0.5% of the training data for
Stable Diffusion V1.1, i.e. 12 million of 2.3 billion instances from
LAION 2B [109], artists like Karen Hallion17 18 found out that their
copyrighted images were used as training data without their con-
sent [11]. And as noted in Section 3, image generators like Stable
Diffusion have been shown to memorize images, outputting replicas
of iconic photographs and paintings by artists [27, 92].
This type of digital forgery causes a number of harms to artists,
many of whom are already struggling to support themselves and can
only perform their artistic work while having other “day” jobs [70].
First, as discussed in Section 4.1, using artists’ works without com-
pensation adds to the already precarious positions that the majority
of professional artists are in [70, 92, 138]. In addition to the lack of
compensation, using artists’ works without their consent can cause
them reputational damage and trauma. Users of image generated
art can mimic an artist’s style by finetuning models like Stable
Diffusion on specific artists’ images, with companies like Wombo
even offering services to generate art in the style tied to specific
groups of artists like Studio Ghibli [133]. A number of artists have
described this practice as “invasive” and noted the manner in which
it causes them reputational damage. After a Reddit user posted
images generated using artist Hollie Mengert’s name as a prompt,
Mengert mentioned that “it felt invasive that my name was on this
tool, I didn’t know anything about it and wasn’t asked about it.”19
She further noted her frustration with having her name associated
with images that do not represent her style except at “the most
surface-level.”
This type of invasive style mimicry can have more severe con-
sequences if an artist’s style is mimicked for nefarious purposes
such as harassment, hate speech and genocide denial. In her New
York Times Op-ed [8], artist Sarah Andersen writes about how even
before the advent of image generators people edited her work "to
reflect violently racist messages advocating genocide and Holocaust
denial, complete with swastikas and the introduction of people get-
ting pushed into ovens. The images proliferated online, with sites
like Twitter and Reddit rarely taking them down." She adds that
"Through the bombardment of my social media with these images,
the alt-right created a shadow version of me, a version that advo-
cated neo-Nazi ideology. . . I received outraged messages and had to
contact my publisher to make my stance against this ultraclear.” She
underscores how this issue is exacerbated by the advent of image
generators, writing "The notion that someone could type my name
into a generator and produce an image in my style immediately
disturbed me. . . I felt violated” [8]. As we discussed in Section 3, an
16 https://laion-aesthetic.datasette.io/laion-aesthetic-6pls/images
17 https://web.archive.org/web/20230811043246/https://twitter.com/Khallion/status/
1615464905565429760
18 https://web.archive.org/web/20230117153958/https://twitter.com/shoomlah/status/
1615215285526757381
19 https://waxy.org/2022/11/invasive-diffusion-how-one-unwilling-illustrator-
found-herself-turned-into-an-ai-model/
AIES ’23, August 08–10, 2023, Montréal, QC, Canada
artist’s style is their unique voice, formed through their life experi-
ences. Echoing Hollie Mengert’s point about the invasive nature
of style mimicry, Andersen adds: “The way I draw is the complex
culmination of my education, the comics I devoured as a child and
the many small choices that make up the sum of my life. The details
are often more personal than people realise.” Thus, tools trained on
artists’ works and which allow users to mimic their style without
their consent or compensation, can cause significant reputational
damage by impersonating artists and spreading messages that they
do not endorse.
4.3 Hegemonic Views and Stereotyping
Beyond the appropriation of individual identities, image generators
have been shown to appropriate and distort identities of groups,
encode biases, and reinforce stereotypes [87, 98, 119]. Introducing
In/Visible, an exhibition exploring the intersection of AI and art,
Senegalese artist Linda Dounia Rebeiz writes: “Any Black person
using AI today can confidently attest that it doesn’t actually know
them, that its conceptualization of their reality is a fragmentary,
perhaps even violent, picture. . . Black people are accustomed to
being unseen. When we are seen, we are accustomed to being mis-
represented. Too often, we have seen our realities ignored, distorted,
or fabricated. These warped realities, often political instruments
of exclusion, follow us around like shadows that we can never
quite shake off” [64]. In an interview, the artist gives examples of
stereotypes perpetuated through image generators. For instance,
she notes that the images generated by Dall-E 2 pertaining to her
hometown Dakar were wildly inaccurate, depicting ruins and desert
instead of a growing coastal city [114]. Similarly, US-based artist
Stephanie Dinkins discusses encountering significant distortions
when prompting image generators to generate images of Black
women [114].
There are already cases of people producing images embody-
ing their view of other populations. In a 2018 New Yorker article,
Lauren Michelle Jackson writes about a white British photogra-
pher, Cameron-James Wilson, who created a dark skinned synthetic
model which he called “Shudu Gram,” and the “World’s first Digital
Supermodel” [77]. The synthetic model, which he created using
a free 3D modeling software called DAZ3D20 , first appeared on
Instagram wearing “iindzila, the neck rings associated with the
Ndebele people of South Africa” [77]. Jackson licensed the image
to various entities such as Balmain21 and Ellesse, 22 many of whom
were criticized for their lack of diversity in hiring [96]. Now, with-
out compensation to any Ndebele people, magazines like Vogue23
profit off of an idealized conception of someone from that commu-
nity, imagined in the mind of a white man who is compensated for
creating that image. Writer Francesca Sobande writes that this is
another iteration of “the objectification of Black people, and the
commodification of Blackness” [115]. Five years later, on March 6
20 https://www.daz3d.com/
21 https://projects.balmain.com/us/balmain/balmains-new-virtual-army
22 https://hypebae.com/2019/2/ellesse-ss19-campaign-shudu-virtual-cgi-digital-
influencer-model
23 https://www.vogue.com.au/fashion/trends/meet-shudu-the-digital-supermodel-
who-is-changing-the-face-of-fashion-one-campaign-at-a-time/news-story/80a96d
3d70043ed2629b5c0bc03701c1
368
2023, entrepreneur Danny Postma announced the launch of a com-
pany, Deep Agency, that rents image generated synthetic models
as a service24 , making the type of practice described by Jackson
more likely to occur at scale.
Due to these questions of who gets to use (and profit from) these
tools by representing which cultures in what way, participants from
Pakistan, India and Bangladesh surveyed in [98] raised “concerns
about artist attribution, commodification, and the consequences
of separating certain art forms from their traditional roots,” with
some questioning which cultural products should be included in
the training set of image generators. To expose these issues, Quadri
et al. recommend further examination of the cultural harms posed
by image generators, including perpetuating cultural hegemony,
erasure or stereotyping [98].
4.4 Chilling Effects on Cultural Production and
Consumption
The harms discussed in the prior sections have created a chilling
effect among artists, who, as artist Steven Zapata notes, are al-
ready a traumatized community with many members struggling to
make ends meet [137]. First, students who foresee image generators
replacing artists have become demoralized and dissuaded from hon-
ing their craft and developing their style [138]. Second, both new
and current artists are becoming increasingly reluctant to share
their works and perspectives, in an attempt to protect themselves
from the mass scraping and training of their life’s works [92, 138].
Independent artists today share their work on social media plat-
forms and crowdfunding campaigns, and sell tutorials, tools, and
resources to other artists on various sites or at art-centric trade
shows 25 . For most artists, gaining enough visibility on any of these
platforms (online or in person) is extremely competitive, taking
them years to build an audience and fanbase to sell their work and
eventually have the ability to support themselves 26 . Thus, having
less visibility in an attempt to protect themselves from unethical
practices by corporations profiting from their work, further reduces
their ability to receive compensation for their work.
Artists’ reluctance to share their work and teach others also
reduces the ability of prospective artists to learn from experienced
ones, limiting the creativity of humans as a whole. Similar to the
feedback loop created by next generations of large language models
trained on the outputs of previous ones [18], if we, as humanity,
rely solely on AI-generated works to provide us with the media we
consume, the words we read, the art we see, we would be heading
towards an ouroboros where nothing new is truly created, a stale
perpetuation of the past. In [18], the authors warn against a similar
issue with future generations of large language models trained on
outputs of prior ones, and static data that does not reflect social
change.
In his 1916 book titled Art, Clive Bell writes “The starting-point
for all systems of aesthetics must be the personal experience of a
peculiar emotion. The objects that provoke this emotion we call
works of art” [15]. As Steven Zapata notes, we need to “protect
24 https://www.deepagency.com/
25 https://www.muddycolors.com/2019/09/results- of - the- artist- income- goals-
survey-2019/
26 https://news.artnet.com/art-world/artist-financial-stability-survey-1300895
AI Art and its Impact on Artists
the creative human spirit. . . Making art is one of the best ways to
investigate one of the ways you are influenced, and the way to send
how you’re influenced to other people. If we don’t curb this, this
influence can come from AI, AI that can’t discern boundaries, and
influence feelings. Let’s not let it happen” [137].
5 AI ART AND US COPYRIGHT LAW
Given the speed at which image generators have been adopted
and their impact, countries around the world are grappling with
what policies to enact in response. In particular, there is a lot of
uncertainty about whether using copyrighted materials to train
image generators is copyright infringement. Some governmental
bodies, like the EU, will require companies to “document and make
publicly available a summary of the use of training data protected
under copyright law”27 [44], which could trigger copyright lawsuits
if it becomes possible to identify specific instances of copyright
infringement [72]. However, it is not clear what the scope of this
law is and if it requires an itemized list of what is included in the
training data, or only a summary of other key information.
While a number of artists have filed class action lawsuits in
the US against companies providing commercial image genera-
tion tools [129], image generators represent a dynamic between
artists and large-scale companies appropriating their work that has
previously not been examined in US copyright law [56]. This is
due to the unprecedented scale at which artists’ works are being
used to create image generators, the recent proliferation of publicly
available image generators trained on that content, and the level
to which the output of the image generators threatens to displace
artists. Furthermore, this dynamic is distinct because of the data
collection practices by which image generators are developed in
the first place [67].
While some of the harms discussed in Section 4 overlap with
the rights protected by US copyright law, others are not. There
are also a number of unanswered legal questions when it comes
to determining the ways in which copyright law applies to image
generators and both the inputs and outputs that go into creating
these tools. Hence, US copyright law is largely unequipped to tackle
many of the types of harms posed by these systems to content
creators. This lack of certainty about whether copyright applies
means that the companies producing these tools can do so largely
without accountability, unless they are sued for specific violations
of copyright law. And waiting for court determinations on their
lawsuits means that artists may not be able to get recourse until
the cases are resolved. In this section, we highlight specific parts of
US copyright law that may be a source of uncertainty and tension
for artists and companies using their work. We conclude that there
are gaps in the law that do not take into account the social and
economic harm to artists.
5.1 Authorship
Thus far, no works created by an image generator have been given
copyright protection, and authorship is limited to human creators.
The US Copyright Office recently affirmed this position by declining
to recognize the copyrightability of works that were created by an
image generator [90]. In the US, the mere effort required to create
27 https://www.euaiact.com/
369
AIES ’23, August 08–10, 2023, Montréal, QC, Canada
a piece of art work does not, on its own, render the resulting work
protected by copyright law, meaning that the number of prompts
or hours poured into the creation of an image using text-to-image
generators will not on its own qualify the work as copyrightable.28
Moreover, the prompts themselves may be protectable if they are in-
dependently creative, and the resulting work may be copyrightable
if the prompts were part of an active process by which the human
creator exercised judgment by selecting, arranging, or designing
the work 29 . US law also requires that the creator of the work be
the source of the creativity and inventiveness of the work, and the
Copyright Office noted that image generators produce images in an
“unpredictable” way [90] and thus cannot be considered creative or
inventive.
These dimensions of what it means to be an “author” under
copyright law as well as how the law understands the process of
creativity means that image generators on their own cannot create
copyright protected works. How artists interact with these tools
would determine the legal status of the output they create. Given
this uncertainty about the legal status of the image generators’
outputs, we can direct our policy attention to the inputs that go
into creating the tools. There is an opportunity here to exercise
more caution in the ex-ante processes of the tools’ development
such that the artists whose works are used to create the tools are
not harmed, which we discuss in Section 7.
5.2 Fair Use
Fair use is a doctrine in copyright law that permits the unauthorized
or unlicensed use of copyrighted works; whether it is to make copies,
to distribute, or to create derivative works. Whether something
constitutes fair use is determined on a case-by-case basis and the
analysis is structured around four factors30 . Most relevant for artists
and generative AI systems are factors 1 and 4, which look at the
purpose or character of the use and its impact on the market [14].
Part of the first factor includes the question of whether the use is
commercial and “transformative”. Commercial use usually weighs
against finding fair use. If the use is found to be transformative,
however, it can be considered fair use even for commercial purposes,
but not always31 . This is in part due to the fourth factor, which
examines whether a use is a threat to the market of the original
creator’s work.
The question of fair use arises at two points within the image
generation ecosystem. First is when the images used to train the
datasets are copyrighted, and especially if the copyright holders are
small-scale artists. These small scale artists could have an interest
in not allowing their work to be used to create synthetic images,
not only because image generators could be used to produce works
resembling theirs, but because of issues around consent and misuse
of their works for harassment, disinformation and hate speech as
described in Section 4.2. Artists may not want to participate in
the creation of an infrastructure that facilitates other informational
harms, even if the image generator is not creating works resembling
28 Bellsouth Advertising & Pub. v. Donnelley Inf. Pub., 933 F. 2d 952 - Court of Appeals,
11th Circuit 1991
29 Feist Publications, Inc. v. Rural Tel. Serv. Co., 499 U.S. 340 (1991)
30 https://www.govinf o.gov/app/details/USCODE- 2011- title17/USCODE- 2011-
title17-chap1-sec107
31 Fox News Network, LLC v. TVEyes, Inc., No. 15-3885 (2d Cir. 2018)
AIES ’23, August 08–10, 2023, Montréal, QC, Canada
theirs. In this case, their concerns wouldn’t be ones that can be
addressed through fair use.
Moreover, small-scale artists may not want to pursue copyright
infringement claims because of a lack of resources to participate in
prolonged legal battles against powerful companies that claim such
copying is fair use. This again means that copyright law is not the
most effective recourse for them and the question of fair use may
not be addressed in a case involving small-scale artists. The com-
plaint32 filed by Getty Images (a large and well-resourced copyright
holder) against Stability AI is illustrative of the resources needed
to assert copyright claims against companies producing image gen-
erators and the power differential that exists between small-scale
copyright holders and these companies. Copyright infringement
is a concern for small-scale artists but the overall system of how
image generators normalize appropriation of art at the input stages
is a problem that is beyond the scope of fair use considerations.
The second point where fair use is a question is if an image
generator is used to create works that are similar to a human artist,
and as a result compete with the human artist’s market, as we
describe in Section 4.1. In such instances, the fourth factor may
weigh heavily against finding fair use. If the work is being used
in ways that displaces the artist’s market share, or prevents them
from receiving appropriate attribution and compensation, there is
a clear harm in place, and may be addressed through copyright law.
5.3 Derivative Works and Moral Rights
When companies design their products around “mimicking” the
style of an artist, then it becomes difficult to justify the company’s
use as fair use [49]. In such instances, there is a clear connection
between the company’s product and the intended outcome being
harm to the market for the original artist’s work.
Such mimicking or use of an artist’s work and style may also
be covered under moral rights in copyright law. Moral rights vest
in “visual art”, such as paintings and photographs, and protect the
creator’s personal and reputational interest in their work by pre-
venting the distortion or defacement of the original work [80]. The
scope of moral rights in US copyright law is narrowly constructed
for various policy reasons [89], but this area of copyright law may
need more attention as artists try to articulate the harms they face.
6 SHORTCOMINGS OF THE AI RESEARCH
COMMUNITY
In the previous section, we provided a brief analysis of US copyright
law that may be relevant to artists’ fight against the harms they
face due to the proliferation of image generators. In this section, we
discuss how academic researchers’ partnerships with corporations
help the latter sidestep some of these laws aimed at protecting cre-
ators. In her paper titled The Steep Cost of Capture, whistleblower
Meredith Whittaker writes about the level to which academic AI
research has been captured by corporate interests [132]. In The
Grey Hoodie Project: Big tobacco, big tech, and the threat on aca-
demic integrity, Mohamed and Moustafa Abdalla liken this capture
to the tobacco and fossil fuel industries, noting that corporations
fund academics aligned with their goals, the same way that tobacco
32 https://news.bloomberglaw.com/ip-law/getty-images-sues-stability-ai-over-art-
generator-ip-violations
370
companies funded doctors that claimed that cigarettes did not cause
cancer [1]. In her article, “You are not a stochastic parrot,” Liz Weil
notes “The membrane between academia and industry is permeable
almost everywhere; the membrane is practically nonexistent at
Stanford, a school so entangled with tech that it can be hard to tell
where the university ends and the businesses begin” [131]. This
corporate entanglement means that the academic research agenda
is increasingly being set by researchers who align themselves with
powerful corporate interests [51, 52, 132].
6.1 Data Laundering
One of the results of this corporate academic partnership has been
data laundering [34]. Similar to money laundering, where business
fronts are created to move money around while obfuscating the
source of illicit funds, researchers have argued that companies use
data laundering to obtain data through nonprofits that are then
used in for profit organizations [10].
The LAION dataset used to train Stability AI’s Stable Diffusion
model, which is also used in their commercial Dream Studio product,
is one such example33 . While LAION is a nonprofit organization,
the paper announcing the LAION-5B dataset notes that Stability AI
CEO Emad Mostaque “provided financial and computation support
for open source datasets and models” [109]. The dataset’s associated
datasheet further answers the question “Who funded the creation of
this dataset” with “This work was sponsored by Hugging Face and
Stability AI.” As we mentioned in Section 6, while US copyright law
is not fully equipped to resolve disputes related to image generated
content, companies are more likely to be granted fair use exceptions
in US copyright law if they claim that the dataset was gathered
for research purposes, even if they end up using it for commercial
products. According to the US copyright office, “Courts look at
how the party claiming fair use is using the copyrighted work, and
are more likely to find that nonprofit educational and noncommer-
cial uses are fair.”34 This allows corporations like Stability AI to
raise $101M in funding with a $1B valuation35 , using datasets that
contain artists’ works without their consent or attribution. The
accountability for the dataset creation and maintenance, on the
other hand, including copyright or privacy issues, is shifted to the
nonprofit that collected it. Thus, while there is no legal distinction
at present between data laundering and the normative data mining
practices in the machine learning communities, this question needs
more attention when the issue of fair use discussed in Section 5.2
arises in the context of image generators.
6.2 Power, ML Fairness, and AI Ethics
In the Moral Character of Cryptographic Work, cryptographer
Philip Rogaway notes that the cryptographic community bears
the responsibility of failing to stop the rise of surveillance [105].
One of the main reasons for this disconnect, according to him, is
that cryptographers fail to take into account how power affects
their analyses, and have a “politically detached posture,” writing
“if power is anywhere in the picture, it is in the abstract capacities
33 https://stability.ai/blog/stablediffusion2-1-release7-dec-2022
34 https://www.copyright.gov/fair-use/
35 https://techcrunch.com/2022/10/17/stability- ai- the- startup- behind- stable-
diffusion-raises-101m/
AI Art and its Impact on Artists AIES ’23, August 08–10, 2023, Montréal, QC, Canada

of notional adversaries or, in a different branch of our field, the Arte es Ética suggests that each image carry “a digital signature”
power expenditure, measured in watts, for some hardware.” Ex- in its metadata, which is disclosed along with the generated image.
cept for a few exceptions, the machine learning fairness and AI Regulation that mandates that organizations disclose their train-
ethics communities have similarly failed to stop the harms caused ing data, at the very least to specific bodies that can verify that
by image generators proliferated by powerful entities, due to their people’s images were not used without their consent, is needed
disproportionate focus on abstract concepts like defining fairness in order to enforce the opt-in requirements artists are demanding.
metrics [84, 110, 136], rather than preventing harm to various com- Such a mandate will likely exist outside of conventional copyright
munities. We urge the machine learning and AI ethics research requirements. However, algorithmic accountability regimes and
communities to orient their focus towards preventing and mitigat- recently proposed laws like the Algorithmic Accountability Act of
ing harms caused to marginalized communities, in order to prevent 2022 in the US [111], or the transparency requirements of the EU’s
further casualties of which the art community is only one. AI Act that would require datasheets [54] or similar data documen-
tation [44], may be preliminarily useful in instituting disclosure
7 RECOMMENDATIONS TO PROTECT requirements for companies.
ARTISTS However, most of these existing measures require individuals
to prove harm, rather than placing the onus on organizations to
To fight back against the harms that artists have already experi-
show lack of harm before proliferating their products. There need
enced, they have filed class action lawsuits in the US against Mid-
to be pathways toward better accountability of the entities and
journey, Stable Diffusion and DeviantArt [129], organized protests,
stakeholders that create the image generators in the first place,
boycotted online services like ArtStation that allowed image gener-
rather than placing additional burdens on artists to prove that they
ated content on their platforms36 , and continue to raise aware-
have been harmed. While auditing, reporting, and transparency are
ness about the impact of image generators on their communi-
well-known possible proposals for regulating AI in general [17, 22,
ties [92, 137, 138]. However, as discussed in Section 5, the US courts
54, 83, 100], formulating sector and industry specific proposals is
can take years to issue a decision, during which more artists would
essential when it comes to effective governance [100], and is what
be harmed, and current US copyright law is ill equipped to protect
will be needed for image generators and art.
artists. Because of this, artists themselves have suggested a number
Regulation, even if successfully passed, takes a long time to be
of regulations to protect them.
enforced however, and is by its very nature reactive. As artist Steven
A letter to members of The Costume Designers Guild, Local
Zapata asks: “What are we going to do. . . to prevent this recurring
892, a union of professional costume designers, assistant costume
over and over again” [137]? This is a fundamental question that re-
designers, and illustrators working in film, television, commercials
quires us to understand why we are in a position where prominent
and other media37 , suggests legislation to allow “using AI derived
machine learning researchers have used their skills to disenfran-
imagery strictly for reference purposes and making it unacceptable
chise artists. One answer is the corporate capture of AI research
to hand over a fully AI generated work as a finished concept” [126].
that we discussed in Section 6.1. To combat this capture, computer
Visual artists who paint in a more representational style usually
scientist Timnit Gebru suggests having government research fund-
work from photo reference or build sculptures to understand how
ing that is not tied to the military, in order to have “alternatives to
lighting works, for example, using stock/licensed photography and
the hugely concentrated power of a few large tech companies and
assets, or the artist’s own work 38 . This would allow artists to use
the elite universities closely intertwined with them” [51].
image generators to provide inspiration in the way that nature,
A few researchers in machine learning have come to the defense
for example, is a source of inspiration to many artists. The art
of artists but they are much smaller in number than those working
collective Arte es Ética suggests having a metric to quantify the
on image generators without attempting to mitigate their harms.
amount of human interaction with an image generator to determine
For instance, University of Chicago student Shawn Shan and his col-
whether or not a generated image is copyrightable, with a 25% or
laborators, advised by security professor Ben Y. Zhao, created a tool
less interaction level being uncopyrightable [41].
called Glaze that allows artists to add perturbations to their images
While these proposals may address the issue of economic loss,
which would prevent diffusion model based generators from being
they do not stop the use of artists’ work for training image genera-
used to mimic their styles [112]. The researchers collaborated with
tors without their consent or compensation. Additional proposed
1000 artists, going to town halls and creating surveys to understand
regulations by Arte es Ética address this issue by recommending
their concerns. While building Glaze, Shawn Shan et al. measured
legislation that requires the explicit consent of content creators
their success by how much the tool was addressing the artists’ con-
before their material is used for generative AI models [41]. In order
cerns. This is an example of research that is conducted in service
to do this, they suggest having detection and filtering algorithms
of specific groups, using a process that identifies stakeholders and
to ensure that uploaded content belongs to creators who have con-
values that should be incorporated in the work, rather than the
sented to their work being licensed or opted-in for use as training
current trend of claiming to build models with “general” capabilities
data. Similar to [18]’s recommendations to ensure that synthetic
that do not perform specific tasks in well defined domains [53, 101].
texts generated by LLMs be “watermarked and thus detectable,”
We echo [18]’s recommendations to use methodologies like value
sensitive design and design justice [33, 48] to identify stakeholders
36 https://www.theverge.com/2022/12/23/23523864/artstation-removing-anti-ai-
and their values, and work on systems that meaningfully incorpo-
protest-artwork-censorship
37 https://www.costumedesignersguild.com/ rate them. These processes encourage researchers and practitioners
38 https://cynthia-sheppard.squarespace.com/#/burn-out/ to consult with visual artists and build tools that make their lives

371
AIES ’23, August 08–10, 2023, Montréal, QC, Canada
easier, rather than claiming to create tools that "democratize art"
without consulting them, as a number of artists have noted [40, 60].
In summary, we advocate for regulation that prevents organiza-
tions from using people’s content to train image generators without
their consent, funding for AI research that is not entangled with
corporate interests, and task specific works in well defined domains
that serve specific communities. It is much easier to accomplish
these goals if machine learning researchers are trained in a manner
that helps them understand how technology interacts with power,
rather than the “view from nowhere” stance that has been critiqued
by feminist scholars, which teaches scientists and engineers that
their work is neutral [50, 105]. We thus advocate for a computer
science education system that stresses the manner in which power
interacts with technology [19, 105].
8 CONCLUSION
In this paper, we have reviewed the chilling impact of image gener-
ators on the art community, ranging from economic loss, to reputa-
tional damage and stereotyping. We summarized recommendations
to protect artists, including new regulation that prohibits training
image generators on artists’ works without opt-in consent, and spe-
cific tools that help artists protect against style mimicry. Our work
is rooted in our argument that art is a uniquely human endeavor.
And we question who its further commodification will benefit. As
artist Steven Zapata asks, “How can we get clear on the things we
do not want to forfeit to automation?” [137]
Image generators can still be a medium of artistic expression
when their training data is not created from artists’ unpaid labor,
their proliferation is not meant to supplant humans, and when
the speed of content creation is not what is prioritized. One such
example is the work of artist Anna Ridler, who created a piece called
Mosaic Virus in 201939 , generating her own training data by taking
photos of 10, 000 Tulips, which itself is a work of art she titled
Myriad (Tulips). She then trained a GAN based image generator
with this data, creating a video where the appearance of a tulip is
controlled by the price of bitcoin, “becoming more striped as the
price of bitcoin goes up—it was these same coveted stripes that
once triggered tulip mania...a 17th-century phenomenon which
saw the price of tulip bulbs rise and crash...It is often held up as
one of the first recorded instances of a speculative bubble" [21]. If
we orient the goal of image generation tools to enhance human
creativity rather than attempt to supplant it, we can have works of
art like those of Anna Ridler that explore its use as a new medium,
and not those that appropriate artists’ work without their consent
or compensation.
ACKNOWLEDGMENTS
Thank you to Emily Denton, Énora Mercier, Jessie Lam, Karla Ortiz,
Neil Turkewitz, Steven Zapata, and the anonymous peer reviewers
for giving us invaluable feedback.
REFERENCES
[1] Mohamed Abdalla and Moustafa Abdalla. 2021. The Grey Hoodie Project: Big
Tobacco, Big Tech, and the Threat on Academic Integrity. In Proceedings of
the 2021 AAAI/ACM Conference on AI, Ethics, and Society (Virtual Event, USA)
39 http://annaridler.com/mosaic-virus
372
(AIES ’21). Association for Computing Machinery, New York, NY, USA, 287–297.
https://doi.org/10.1145/3461702.3462563
[2] Adobe. 2022. Generative AI Content. Retrieved March 13, 2023 from https:
//helpx.adobe.com/ca/stock/contributor/help/generative-ai-content.html
[3] Adobe. 2023. Generative AI for creatives - Adobe Firefly. Retrieved July 6, 2023
from https://www.adobe.com/sensei/generative-ai/firefly.html
[4] Stability AI. 2022. Stable Diffusion Launch Announcement. Retrieved March
13, 2023 from https://stability.ai/blog/stable-diffusion-announcement
[5] Stability AI. 2022. Stable Diffusion Public Release. https://stability.ai/blog/stable-
diffusion-public-release/
[6] Thomas M Alexander. 2013. The human eros: Eco-ontology and the aesthetics of
existence. Fordham University Press.
[7] Abdullah Alfaraj. 2023. Auto-Photoshop-StableDiffusion-Plugin. Retrieved
March 13, 2023 from https://github.com/AbdullahAlfaraj/Auto-Photoshop-
StableDiffusion-Plugin
[8] Sarah Andersen. 2022. The Alt-Right Manipulated My Comic. Then A.I. Claimed
It. New York Times (Dec. 31, 2022). https://www.nytimes.com/2022/12/31/opini
on/sarah-andersen-how-algorithim-took-my-work.html
[9] Artbreeder. 2022. artbreeder. Retrieved March 13, 2023 from https://www.artb
reeder.com/
[10] Andy Baio. 2021. AI Data Laundering: How Academic and Nonprofit Researchers
Shield Tech Companies from Accountability. Waxy (2021).
[11] Andy Baio. 2022. Exploring 12 Million of the 2.3 Billion Images Used to Train
Stable Diffusion’s Image Generator. Retrieved July 6, 2023 from https://wa
xy.org/2022/08/exploring-12-million-of -the-images-used-to-train-stable-
diffusions-image-generator/
[12] Ollie Barder. 2017. Katsuhiro Otomo On Creating ’Akira’ And Designing The
Coolest Bike In All Of Manga And Anime. Forbes (2017).
[13] Alexis T Baria and Keith Cross. 2021. The brain is a computer is a brain:
neuroscience’s internal debate and the social significance of the Computational
Metaphor. arXiv preprint arXiv:2107.14042 (2021).
[14] Barton Beebe. 2008. An Empirical Study of U.S. Copyright Fair Use Opinions,
1978-2005. University of Pennsylvania Law Review 156, 3 (2008). https://schola
rship.law.upenn.edu/penn_law_review/vol156/iss3/1/
[15] Clive Bell. 1916. Art. Chatto & Windus.
[16] Emily M Bender. 2022. Resisting Dehumanization in the Age of “AI”. (2022).
[17] Emily M Bender and Batya Friedman. 2018. Data statements for natural lan-
guage processing: Toward mitigating system bias and enabling better science.
Transactions of the Association for Computational Linguistics 6 (2018), 587–604.
[18] Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret
Shmitchell. 2021. On the Dangers of Stochastic Parrots: Can Language Models
Be Too Big?. In Proceedings of the 2021 ACM Conference on Fairness, Account-
ability, and Transparency (Virtual Event, Canada) (FAccT ’21). Association for
Computing Machinery, New York, NY, USA, 610–623. https://doi.org/10.1145/
3442188.3445922
[19] Ruha Benjamin. 2020. Race after technology: Abolitionist tools for the new jim
code.
[20] blueturtleai. 2023. gimp-stable-diffusion. Retrieved March 13, 2023 from
https://github.com/blueturtleai/gimp-stable-diffusion
[21] Ruby Boddington. 2019. Anna Ridler uses AI to turn 10,000 tulips into a video
controlled by bitcoin. It’s Nice That (2019).
[22] Karen L Boyd. 2021. Datasheets for datasets help ml engineers notice and
understand ethical issues in training data. Proceedings of the ACM on Human-
Computer Interaction 5, CSCW2 (2021), 1–27.
[23] Jo Ann Boydston. 2008. The Middle Works of John Dewey, Volume 9, 1899-1924:
Democracy and Education 1916. Southern Illinois University Press.
[24] Andrew Brock, Jeff Donahue, and Karen Simonyan. 2018. Large Scale GAN
Training for High Fidelity Natural Image Synthesis. https://doi.org/10.48550/A
RXIV.1809.11096
[25] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan,
Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda
Askell, et al. 2020. Language models are few-shot learners. Advances in neural
information processing systems 33 (2020), 1877–1901.
[26] Canva. 2022. Text to Image - Canva Apps. Retrieved March 13, 2023 from
https://www.canva.com/apps/text-to-image
[27] Nicholas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag,
Florian Tramèr, Borja Balle, Daphne Ippolito, and Eric Wallace. 2023. Extracting
Training Data from Diffusion Models. https://doi.org/10.48550/ARXIV.2301.13
188
[28] Electronic Privacy Information Center. 2023. Generating Harms. Generative
AI’s Impact & Paths Forward. Retrieved July 6, 2023 from https://epic.org/doc
uments/generating-harms-generative-ais-impact-paths-forward/
[29] Chen Chen, Jie Fu, and Lingjuan Lyu. 2023. A Pathway Towards Responsible
AI Generated Content. https://doi.org/10.48550/ARXIV.2303.01325
[30] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan,
and Ilya Sutskever. 2020. Generative Pretraining From Pixels. In Proceedings of
the 37th International Conference on Machine Learning (Proceedings of Machine
AI Art and its Impact on Artists
Learning Research, Vol. 119), Hal Daumé III and Aarti Singh (Eds.). PMLR, 1691–
1703. https://proceedings.mlr.press/v119/chen20s.html
[31] Neil Clarke. 2023. A Concerning Trend. (2023).
[32] Samantha Cole. 2023. Netflix Made an Anime Using AI Due to a ’Labor Shortage,’
and Fans Are Pissed. Vice (2023).
[33] Sasha Costanza-Chock. 2020. Design justice: Community-led practices to build
the worlds we need. The MIT Press.
[34] Subhasish Dasgupta. 2005. Encyclopedia of virtual communities and technologies.
IGI Global.
[35] Li Deng. 2012. The mnist database of handwritten digit images for machine
learning research. IEEE Signal Processing Magazine 29, 6 (2012), 141–142.
[36] John Dewey. 1934. Art as Experience. Penguin Group (USA) Inc.
[37] Prafulla Dhariwal and Alex Nichol. 2021. Diffusion Models Beat GANs on Image
Synthesis. https://doi.org/10.48550/ARXIV.2105.05233
[38] Alexei A. Efros and William T. Freeman. 2001. Image Quilting for Texture
Synthesis and Transfer. In Proceedings of the 28th Annual Conference on Computer
Graphics and Interactive Techniques (SIGGRAPH ’01). Association for Computing
Machinery, New York, NY, USA, 341–346. https://doi.org/10.1145/383259.383296
[39] Ziv Epstein, Sydney Levine, David G Rand, and Iyad Rahwan. 2020. Who gets
credit for AI-generated art? Iscience 23, 9 (2020).
[40] Arte es Ética. 2023. Arte es Ética presente en la ACM FAccT Conference 2023.
Retrieved July 6, 2023 from https://arteesetica.org/arte-es-etica-presente-en-la-
acm-facct-conference-2023/
[41] Arte es Ética. 2023. Un manifiesto para comprender y regular la I.A. generativa.
(April 2023). https://arteesetica.org
[42] Patrick Esser, Robin Rombach, and Björn Ommer. 2020. Taming Transformers for
High-Resolution Image Synthesis. https://doi.org/10.48550/ARXIV.2012.09841
[43] Alberto Cerdá Estarelles. 2023. ue-stable-diffusion. Retrieved March 13, 2023
from https://github.com/albertotrunk/ue-stable-diffusion
[44] European Commission. 2021. The Artificial Intelligence Act. (2021).
[45] Everimaging. 2022. Fotor. Retrieved March 13, 2023 from https://www.fotor.co
m/features/ai-image-generator/
[46] Creative Fabrica. 2022. CF Spark AI Art Generator. Retrieved March 13, 2023
from https://www.creativefabrica.com/spark/ai-image-generator/
[47] Mureji Fatunde and Crystal Tse. 2022. Stability AI Raises Seed Round at $1 Billion
Value. Bloomberg (Oct. 17, 2022). https://www.bloomberg.com/news/articles/
2022-10-17/digital-media-firm-stability-ai-raises-funds-at-1-billion-value
[48] Batya Friedman. 1996. Value-sensitive design. interactions 3, 6 (1996), 16–23.
[49] Thania Garcia. 2023. David Guetta Replicated Eminem’s Voice in a Song Using
Artificial Intelligence. Variety (Feb. 8, 2023). https://variety.com/2023/music/n
ews/david-guetta-eminem-artificial-intelligence-1235516924/
[50] Timnit Gebru. 2019. Chapter on race and gender. Oxford Handbook on AI Ethics.
Available from: http://arxiv. org/abs/1908.06165 [Accessed 28th April 2020] (2019).
[51] Timnit Gebru. 2021. For truly ethical AI, its research must be independent from
big tech. Guardian (2021).
[52] Timnit Gebru. 2022. Effective Altruism is Pushing a Dangerous Brand of ‘AI
Safety’. Wired (2022).
[53] Timnit Gebru. 2023. Eugenics and The Promise of Utopia through Artificial
General Intelligence.
[54] Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman
Vaughan, Hanna Wallach, Hal Daumé Iii, and Kate Crawford. 2021. Datasheets
for datasets. Commun. ACM 64, 12 (2021), 86–92.
[55] Deep Dream Generator. 2022. Deep Dream Generator. Retrieved March 13,
2023 from https://deepdreamgenerator.com/
[56] Jessica Gillotte. 2020. Copyright Infringement in AI-Generated Artworks. UC
Davis Law Review 53, 5 (2020). https://ssrn.com/abstract=3657423
[57] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. 2016. Deep Learning. MIT
Press. http://www.deeplearningbook.org.
[58] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-
Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative
Adversarial Networks. https://doi.org/10.48550/ARXIV.1406.2661
[59] Abhishek Gupta. 2022. Unstable Diffusion: Ethical challenges and some ways
forward. Retrieved March 13, 2023 from https://montrealethics.ai/unstable-
diffusion-ethical-challenges-and-some-ways-forward/
[60] The Distributed AI Research Institute. 2023. Stochastic Parrots Day: What’s
Next? A Call to Action. Retrieved July 6, 2023 from https://peertube.dair-
institute.org/w/a2KCCzj2nyfrq5FvJbL9jY
[61] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. 2016. Image-to-
Image Translation with Conditional Adversarial Networks. https://doi.org/10
.48550/ARXIV.1611.07004
[62] Philip J Ivanhoe. 2000. Confucian moral self cultivation. Hackett Publishing.
[63] Nina Kalitina and Nathalia Brodskaya. 2012. Claude Monet. Parkstone Interna-
tional.
[64] KALTBLUT. 2023. Feral File presents In/Visible: Where AI Meets Artistic
Diversity, A Fascinating Encounter of Limitations and Storytelling. Retrieved
July 6, 2023 from https://www.kaltblut-magazine.com/f eral-file-presents-
in-visible-where-ai-meets-artistic-diversity-a-fascinating-encounter-of -
373
AIES ’23, August 08–10, 2023, Montréal, QC, Canada
limitations-and-storytelling/
[65] Tero Karras, Samuli Laine, and Timo Aila. 2018. A Style-Based Generator
Architecture for Generative Adversarial Networks. https://doi.org/10.48550/A
RXIV.1812.04948
[66] Kevin Kelly. 2022. Picture Limitless Creativity at Your Fingertips. Wired (2022).
[67] Mehtab Khan and Alex Hanna. 2022. The Subjects and Stages of AI Dataset
Development: A Framework for Dataset Accountability. (Sept. 13, 2022). https:
//doi.org/10.2139/ssrn.4217148
[68] Diederik P Kingma and Max Welling. 2013. Auto-Encoding Variational Bayes.
https://doi.org/10.48550/ARXIV.1312.6114
[69] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. ImageNet Clas-
sification with Deep Convolutional Neural Networks. In Advances in Neural
Information Processing Systems, F. Pereira, C.J. Burges, L. Bottou, and K.Q. Wein-
berger (Eds.), Vol. 25. Curran Associates, Inc. https://proceedings.neurips.cc/p
aper/2012/file/c399862d3b9d6b76c8436e924a68c45b-Paper.pdf
[70] Jennifer C Lena and Danielle J Lindemann. 2014. Who is an artist? New data
for an old question. Poetics 43 (2014), 70–85.
[71] Craiyon LLC. 2022. Craiyon. Retrieved March 13, 2023 from https://www.crai
yon.com/
[72] Freshfields Bruckhaus Deringer LLP. 2023. Has copyright caught up with the
AI Act? Retrieved July 6, 2023 from https://www.lexology.com/library/detail.
aspx?g=d9820844-8983-4aec-88d7-66e385627b4a#:~:text=Among%20the%20a
mendments%20to%20the,used%20for%20training%20those%20models
[73] MyHeritage Ltd. 2022. AI Time Machine. Retrieved March 13, 2023 from
https://www.myheritage.com/ai-time-machine
[74] Michele Marra. 1995. Japanese Aesthetics: The construction of meaning. Philos-
ophy East and West (1995), 367–386.
[75] Michael F Marra. 1999. Modern Japanese aesthetics: a reader. University of
Hawaii Press.
[76] Rick Merritt. 2023. Getty Images bans AI-generated content over fears of legal
challenges. NVIDIA Blog ( 21, 2023). https://blogs.nvidia.com/blog/2023/03/21
/generative-ai-getty-images/
[77] Lauren Michele Jackson. 2018. Shudu Gram Is a White Man’s Digital Projection
of Real-Life Black Womanhood. New Yorker (2018).
[78] Midjourney. 2022. Midjourney. Retrieved March 13, 2023 from https://www.mi
djourney.com/home/
[79] Midjourney. 2023. Version. Retrieved July 7, 2023 from https://docs.midjourne
y.com/docs/model-versions
[80] Martin Miernicki and Irene Ng (Huang Ying). 2021. Artificial intelligence
and moral rights. AI & SOCIETY 36, 1 (March 1, 2021), 319–329. https:
//doi.org/10.1007/s00146-020-01027-6
[81] Zosha Millman. 2023. Yes, Secret Invasion’s opening credits scene is AI-made —
here’s why. Polygon (June 22, 2023). https://www.polygon.com/23767640/ai-
mcu-secret-invasion-opening-credits
[82] Mehdi Mirza and Simon Osindero. 2014. Conditional Generative Adversarial
Nets. https://doi.org/10.48550/ARXIV.1411.1784
[83] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasser-
man, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru.
2019. Model cards for model reporting. In Proceedings of the conference on
fairness, accountability, and transparency. 220–229.
[84] Arvind Narayanan. 21. Fairness definitions and their politics. In Tutorial pre-
sented at the Conf. on Fairness, Accountability, and Transparency.
[85] Andrew Ng and Michael Jordan. 2001. On Discriminative vs. Generative Classi-
fiers: A comparison of logistic regression and naive Bayes. In Advances in Neural
Information Processing Systems, T. Dietterich, S. Becker, and Z. Ghahramani
(Eds.), Vol. 14. MIT Press. https://proceedings.neurips.cc/paper/2001/file/7b7a
53e239400a13bd6be6c91c4f6c4e-Paper.pdf
[86] Alex Nichol and Prafulla Dhariwal. 2021. Improved Denoising Diffusion Proba-
bilistic Models. https://doi.org/10.48550/ARXIV.2102.09672
[87] Leonardo Nicoletti and Dina Bass. 2023. Humans Are Biased. Generative AI Is
Even Worse. Bloomberg (2023). https://www.bloomberg.com/graphics/2023-
generative-ai-bias/
[88] Nkiru Nzegwu. 2005. Art and community: A social conception of beauty and
individuality. A companion to African Philosophy (2005), 415–424.
[89] United States Copyright Office. 2019. Authors, Attribution, and Integrity: Ex-
amining Moral Rights in the United States. https://www.copyright.gov/policy
/moralrights/full-report.pdf
[90] United States Copyright Office. 2023. Re: Zarya of the Dawn (Registration #
VAu001480196). https://www.copyright.gov/docs/zarya-of-the-dawn.pdf
[91] OpenAI. 2022. DALL-E 2. Retrieved March 13, 2023 from https://openai.com/p
roduct/dall-e-2
[92] Karla Ortiz. 2022. Why AI Models are not inspired like humans. (Dec.10 2022).
https://www.kortizblog.com/blog/why-ai-models-are-not-inspired-like-
humans
[93] Stephen Owen. 2020. Readings in Chinese literary thought. Vol. 30. BRILL.
[94] LLC Panabee. 2022. Hotpot. Retrieved March 13, 2023 from https://hotpot.ai/
[95] Javier Portilla and Eero P. Simoncelli. 2000. A Parametric Texture Model Based
on Joint Statistics of Complex Wavelet Coefficients. International Journal of
AIES ’23, August 08–10, 2023, Montréal, QC, Canada

Computer Vision 40, 1 (Oct. 1, 2000), 49–70. https://doi.org/10.1023/A:10265536 [121] Inc. Stablecog. 2022. Stablecog. Retrieved March 13, 2023 from https://stableco
19983 g.com/
[96] Anshuman Prasad, Pushkala Prasad, and Raza Mir. 2011. ‘One mirror in another’: [122] starryai. 2022. starryai. Retrieved March 13, 2023 from https://starryai.com/
Managing diversity and the discourse of fashion. Human Relations 64, 5 (2011), [123] NightCafe Studio. 2022. AI Art Generator. Retrieved March 13, 2023 from
703–724. https://nightcafe.studio/
[97] Inc. Prisma Labs. 2022. Lensa. Retrieved March 13, 2023 from https://prisma- [124] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. 2017.
ai.com/lensa Revisiting Unreasonable Effectiveness of Data in Deep Learning Era. https:
[98] Rida Qadri, Renee Shelby, Cynthia L. Bennett, and Emily Denton. 2023. AI’s //doi.org/10.48550/ARXIV.1707.02968
Regimes of Representation: A Community-centered Study of Text-to-Image [125] Kentaro Takeda, Akira Oikawa, and Yukiko Une. 2023. Generative AI mania
Models in South Asia. In 2023 ACM Conference on Fairness, Accountability, and brings billions of dollars to developers. Nikkei Asia (March 4, 2023). https:
Transparency. ACM. https://doi.org/10.1145/3593013.3594016 //asia.nikkei.com/Spotlight/Datawatch/Generative-AI-mania-brings-billions-
[99] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, of-dollars-to-developers
Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, [126] The Costume Designers Guild, Local 892. 2022. A Letter to Membership. (2022).
Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual [127] Nitasha Tiku. 2022. AI can now create any image in seconds, bringing wonder
Models From Natural Language Supervision. https://doi.org/10.48550/ARXIV and danger. Washington Post (Sept. 28, 2022). https://www.washingtonpost.c
.2103.00020 om/technology/interactive/2022/artificial-intelligence-images-dall-e/
[100] Inioluwa Deborah Raji. 2022. From Algorithmic Audits to Actual Accountability: [128] Unite.AI. 2022. Images.AI. Retrieved March 13, 2023 from https://images.ai/
Overcoming Practical Roadblocks on the Path to Meaningful Audit Interventions [129] James Vincent. 2022. AI art tools Stable Diffusion and Midjourney targeted with
for AI Governance. In Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, copyright lawsuit. The Verge (2022).
and Society. 5–5. [130] James Vincent. 2022. Getty Images bans AI-generated content over fears of
[101] Inioluwa Deborah Raji, Emily M Bender, Amandalynne Paullada, Emily Den- legal challenges. The Verge (Sept. 21, 2022). https://www.theverge.com/2022/9/
ton, and Alex Hanna. 2021. AI and the everything in the whole wide world 21/23364696/getty-images-ai-ban-generated-artwork-illustration-copyright
benchmark. arXiv preprint arXiv:2111.15366 (2021). [131] Liz Weil. 2023. You are not a stochastic parrot. New York Magazine (2023).
[102] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. [132] Meredith Whittaker. 2021. The steep cost of capture. Interactions 28, 6 (2021),
2022. Hierarchical Text-Conditional Image Generation with CLIP Latents. https: 50–55.
//doi.org/10.48550/ARXIV.2204.06125 [133] WOMBO. 2022. Dream by WOMBO. Retrieved March 13, 2023 from https:
[103] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Rad- //dream.ai/
ford, Mark Chen, and Ilya Sutskever. 2021. Zero-Shot Text-to-Image Generation. [134] Writesonic. 2022. Photosonic. Retrieved March 13, 2023 from https://photoson
https://doi.org/10.48550/ARXIV.2102.12092 ic.writesonic.com/
[104] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. 2019. Generating Diverse [135] Xinchen Yan, Jimei Yang, Kihyuk Sohn, and Honglak Lee. 2015. Attribute2Image:
High-Fidelity Images with VQ-VAE-2. https://doi.org/10.48550/ARXIV.1906.00 Conditional Image Generation from Visual Attributes. https://doi.org/10.48550
446 /ARXIV.1512.00570
[105] Phillip Rogaway. 2015. The moral character of cryptographic work. Cryptology [136] Meg Young, Michael Katell, and PM Krafft. 2022. Confronting Power and
ePrint Archive (2015). Corporate Capture at the FAccT Conference. In 2022 ACM Conference on Fairness,
[106] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Accountability, and Transparency. 1375–1386.
Ommer. 2021. High-Resolution Image Synthesis with Latent Diffusion Models. [137] Steven Zapata. 2023. AI/ML Media Advocacy Summit Keynote: Steven Zapata.
https://doi.org/10.48550/ARXIV.2112.10752 https://www.youtube.com/clip/UgkxNwkkN9SnPpSJlnbkDK4q-NgxnaAyPHL
[107] Inc. Runway AI. 2022. Runway. Retrieved March 13, 2023 from https://runway L Presented through the Concept Art Association YouTube Channel.
ml.com/ [138] Steven Zapata. 2023. Who Makes the Art in Utopia? https://www.youtub
[108] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily e.com/clip/UgkxNwkkN9SnPpSJlnbkDK4q- NgxnaAyPHLL Presented at
Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mah- TEDxBerkeley, Berkeley, CA.
davi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and [139] Viola Zhou. 2023. AI is already taking video game illustrators’ jobs in China.
Mohammad Norouzi. 2022. Photorealistic Text-to-Image Diffusion Models with Rest of World (April 11, 2023). https://restofworld.org/2023/ai-image-china-
Deep Language Understanding. https://doi.org/10.48550/ARXIV.2205.11487 video-game-layoffs/
[109] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross
Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell
Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson,
Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. 2022. LAION-5B:
An open large-scale dataset for training next generation image-text models.
https://doi.org/10.48550/ARXIV.2210.08402
[110] Andrew D Selbst, Danah Boyd, Sorelle A Friedler, Suresh Venkatasubramanian,
and Janet Vertesi. 2019. Fairness and abstraction in sociotechnical systems. In
Proceedings of the conference on fairness, accountability, and transparency. 59–68.
[111] Ron Sen. Wyden. 2022. The Algorithmic Accountability Act of 2022. (2022).
[112] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and
Ben Y Zhao. 2023. GLAZE: Protecting Artists from Style Mimicry by Text-to-
Image Models. arXiv preprint arXiv:2302.04222 (2023).
[113] Shutterstock. 2022. Shutterstock. Retrieved March 13, 2023 from https:
//www.shutterstock.com/
[114] Zachary Small. 2023. Black Artists Say A.I. Shows Bias, With Algorithms Erasing
Their History. New York Times (July 4, 2023). https://www.nytimes.com/2023/0
7/04/arts/design/black-artists-bias-ai.html
[115] Francesca Sobande. 2021. CGI influencers are here. Who’s profiting from them
should give you pause. Fast Company (2021).
[116] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli.
2015. Deep Unsupervised Learning using Nonequilibrium Thermodynamics.
https://doi.org/10.48550/ARXIV.1503.03585
[117] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom
Goldstein. 2022. Diffusion Art or Digital Forgery? Investigating Data Replication
in Diffusion Models. https://doi.org/10.48550/ARXIV.2212.03860
[118] Jiaming Song, Chenlin Meng, and Stefano Ermon. 2020. Denoising Diffusion
Implicit Models. https://doi.org/10.48550/ARXIV.2010.02502
[119] Ramya Srinivasan and Kanji Uchino. 2021. Biases in generative art: A causal
look from the lens of art history. In Proceedings of the 2021 ACM Conference on
Fairness, Accountability, and Transparency. 41–51.
[120] A. Srivastava, A.B. Lee, E.P. Simoncelli, and S.-C. Zhu. 2003. On Advances in
Statistical Modeling of Natural Images. Journal of Mathematical Imaging and
Vision 18, 1, 17–33. https://doi.org/10.1023/A:1021889010444

374